ipman-cli 0.3.0

Intelligence Package Manager - Agent skill virtual environment manager

IpMan - Intelligence Package Manager

I can take on ten.

Project-level virtual environments for AI agent skills — like uv/conda, but for skills. One central store, per-skill symlinks, and a lockfile that reproduces your exact skill setup anywhere.

Documentation | 中文文档 | 中文 README

---

The problem: skill bloat

Your agent caps how much context skill listings may use. Install enough global skills and they compete for that budget: descriptions of the skills you actually need get truncated or dropped, irrelevant skills mis-fire, and project A's tooling leaks into project B. Power users already solve this by hand: a central folder of cloned skill repos, symlinked per-project, updated in one place.

IpMan turns that workflow into one command — and adds the piece the manual approach can't give you: reproducibility.

How it works

ip.yaml ──resolve──> central store clone ──pin──> @<commit> worktree
                                                        │
ip.lock <──record── commit + tree hash                  └──symlink── .claude/skills/<name>

• Project-scoped by default. Skills live in the project that needs them. Your agent's context stays lean.
• One source of truth. Each skill repo is cloned once into ~/.ipman/store. Projects link to it; nothing is copied around.
• Pinned and reproducible. ip.lock records the exact commit. Clone a repo on a new machine, run ipman sync, get the identical skill sources back — even if upstream moved on.
• Contribute back. --live mode links a skill's real working copy: fix a bug while you work, commit it, PR it upstream.

ip.lock guarantees skill source reproducibility (MCP servers, plugins and CLI dependencies are out of scope).

Quick start

pip install ipman-cli        # or: uv tool install ipman-cli

cd your-project
ipman init                                                 # creates ip.yaml
ipman add https://github.com/owner/skills-repo#skills/foo  # pick a skill from a repo
ipman list                                                 # what this project uses
ipman doctor                                               # links, drift, health

Commit ip.yaml and ip.lock. Anyone (including future you, on another machine) restores the whole skill environment with:

uvx ipman-cli sync

More commands:

ipman add <path> --adopt      # register your existing ~/GitHub clone, no copying
ipman add <url>#dir --live    # link the live working copy (skill development)
ipman update [name]           # re-pin this project to the source's latest commit
ipman remove <name>           # unlink from this project (store is untouched)
ipman shell init              # optional shell integration

Agent support

Agent	Status
Claude Code (.claude/skills/)	Verified (macOS)
OpenClaw (.openclaw/skills/)	Path mapping, unverified

Adding an agent is a one-line path mapping — the store and lockfile are agent-agnostic by design.

Security note

Third-party skills can instruct your agent to run arbitrary commands. IpMan shows what it installs (source, commit, skill directory) and never installs silently — review before you use. A deeper risk-assessment engine exists in the codebase but is dormant while the core workflow matures.

Status

Early and moving fast. The core loop (init/add/sync/list/remove/update/doctor) is tested on macOS with Claude Code. Windows has a fallback code path but is not yet a supported target. Feedback and issues welcome.

License

Apache License 2.0 — see LICENSE for details.